Advancing Infrastructure for the Era of Agentic AI | Ian Buck at AI Infra Summit 2026

0:00 Today I'm going to talk about advancing infrastructure for the era of agenda AI.

0:04 Basically, you know, the new era of what inference means.

0:09 Uh, some provide some perspective on what it

0:13 looks like from from my desk and also some

0:16 of the innovations that we're working on so that the community

0:19 here can understand what it is that we're building.

0:22 Uh, explain a little bit more about Vera Rubin

0:24 and how things have changed in just a few years, or even just last year.

0:29 Um, and what we're doing about it at a high level,

0:33 I think the universe of computing is enormous.

0:37 Uh, Ian and and Doctor Patterson did explain,

0:41 you know, how computing has changed his perspective,

0:44 going all the way back to magnetic course, to today, to tomorrow.

0:48 Uh, it really shows how infrastructure for computing

0:51 and the ability to compute, transforming data,

0:56 taking information and applying a calculation and producing a result

1:01 and fitting it back around and transforming it again.

1:05 If that's the essence of value creation, storing information is important.

1:09 Communicating information is important,

1:11 but the value gets generated in the computing

1:14 and the transformation of the actual data into a calculation.

1:19 From our standpoint, it's the the platform is enormous.

1:22 I've been doing this for almost been an Nvidia 26 years.

1:27 Um, it is obviously started with, you know,

1:31 bringing a new kind of computing platform online, CUDA.

1:34 And since then we have it's expanded immensely.

1:37 This is a perspective of kind of like the where

1:40 the diversity is of people that are using our GPUs,

1:43 from the hyperscalers to the AI natives, the startups to the major AI labs,

1:49 to the enterprises, as well as the edge to the edge.

1:53 And this is a combination of all things

1:56 from both AI models as well as traditional simulation.

1:59 Or now the combination of both because they actually support each other.

2:04 There are now over, I believe, 3 million models on hugging face.

2:10 About 3000 of them are actually about 90% of the downloads.

2:15 The diversity models is continuously expanding.

2:17 The workloads are expanding from vision to physical sciences to life sciences,

2:21 speech to natural languages to robotics.

2:23 We have AI models for quantum computing and simulation,

2:27 to the multimodal and advanced models

2:29 of which are exploring new kinds of architectures.

2:32 On top of that, there is a whole

2:37 litany of different simulation libraries and applications for computing,

2:43 ranging from computational lithography to key to decision making,

2:48 to medical imaging, to weather analysis to genomics.

2:52 Computing has become a A capability a necessary part of actually science,

2:59 of industry, of consumer engagement, of enterprise calculations.

3:05 And it is spanning both AI and simulation and in calculations.

3:11 We have actually over 8000 major Cuda applications

3:15 that are actually out there running on these platforms,

3:19 and that we have been building that platform for myself about 25 years.

3:24 Um, what does it look like today?

3:26 So this is Vera Ribbon.

3:29 It's designed to be the next generation inference energetic platform.

3:33 Obviously, we continue to support training and doing training at scale,

3:38 which will be a capability we can offer.

3:40 But but what what brings that to market brings us to life is the AI factory.

3:46 And Vera Rubin was designed to be that.

3:48 Um, it has it's built on a rich ecosystem.

3:51 We have over about 10 million developers developing on NVIDIA GPUs.

3:55 All around the world, at least that we can.

3:58 We can track.

3:59 Um, it spans the CSPs, the NCPs, the OEMs, the ODM.

4:03 There's over 300 major companies that that are, uh,

4:07 have built a supply chain that help deliver this platform.

4:12 Uh, it's and it's amazing to watch it happen.

4:15 It's highly fungible.

4:16 And I think that's one of the things that makes it so, um,

4:20 valued and helpful to the world is that you're providing

4:24 a computing platform that can run all the workloads, every model, uh,

4:28 whether it be opened or closed or frontier, it can support the pre-training,

4:33 the post training, the inference, the authentic,

4:36 and of course, the accelerated computing, uh,

4:39 all of which are tapping together at once in the authentic workload.

4:43 Uh, and then, of course, it's durable.

4:46 David talked about the, uh, the GPUs and TC still offering the original Volta's,

4:51 uh, which goes back quite a ways.

4:54 Almost ten.

4:54 You can still find those GPUs that we

4:57 launched ten years ago are providing value.

5:00 In fact, if you can track the dollar per hour instance pricing

5:04 of a A100, which I think we launched right at the beginning of Covid,

5:08 has only increased over time.

5:10 Uh, and continues to be in, in high demand.

5:14 Uh.

5:17 So what's changed on the side?

5:19 Definitely the workload.

5:20 You know, a lot of the benchmarking this is actually, you know,

5:24 chat the original 2023 breakthrough workload, um,

5:27 which became the foundation for benchmarks like inference X.

5:31 Um, it had a one K, you know, number of input tokens.

5:36 Maybe you had a, uh, kV cache size about 4K for the system prompt,

5:40 as well as the other user data that you may have.

5:43 Um, and then you were interacting it with turns.

5:45 So you'd ask a question, it would reply,

5:47 and then you maybe ask a follow on and on average about three turns.

5:51 And it was a human human in the loop method.

5:53 There was no agents back then,

5:55 but this was the foundation for benchmarking and to understanding our value.

5:59 How well could you run this workload,

6:01 whether it be a DeepSeek model or a closed,

6:04 you know, ChatGPT like interface and what the world benchmark back in 2023?

6:09 Today, agent tech is just so different,

6:12 and I can't stress how much it has provided both a challenge.

6:17 It's incredibly demanding.

6:18 It's actually 100 times more demanding in terms of just raw numbers.

6:23 Um, this is actually from the agent X benchmark.

6:26 Um, I'm not going to say it's representative of everybody.

6:30 But, you know, how we benchmark today versus inference before,

6:33 you know, the average ISO is 142,000 tokens.

6:36 Um, and it's dynamic.

6:38 You need to support short ISO one k ten k

6:43 100 k all the way up 142 to 200 and beyond.

6:48 Every aisle is a whole nother set of kernel

6:52 optimizations of tuning and dialing in of all the math,

6:57 and has as a massively increased the complexity of optimizing the software.

7:02 The kV cache size has gotten huge.

7:04 The models support up to a million tokens.

7:07 You may start off small,

7:09 but as the agents are doing multiple turns now, it's not humans and loops,

7:13 but agents are asking over and over and over again that adds to the context.

7:17 Now all this kV has to be managed across an entire data center.

7:21 And then finally the number of turns is exploding.

7:24 AI was the designers of services could rely on the fact

7:28 that you would take time to read the answer and give another one.

7:32 And in terms of scheduling the work, that was relatively easy.

7:35 You've now ages have now taken humans out of the loop,

7:39 and the compute will be consumed as fast as the agent can make the turn,

7:44 which is really fast and has exploded all of these numbers for agent X.

7:49 Actually, they only have about four sub agents in the benchmark.

7:53 We have thousands of agents in commercial services,

7:56 but they have to do intelligent routing.

7:58 The model now has to decide and call

8:01 out all those different sub agents across sub platform.

8:03 And the KV cache needs to be preserved across all of it.

8:06 So if you look at the workload on the bottom, it's it's much more complicated.

8:11 What was previously tokens in Roma model tokens out.

8:16 We now have orchestration CPU in managing

8:19 the dynamic Kubernetes containers across the helm charts.

8:23 We have the context, an LLMs processing the context and observing the data.

8:29 Another level I'm doing reasoning.

8:30 Another role in deciding which sub agents to do

8:33 to switch between the different models in the next turn, or what tools to call.

8:38 Finally, we have all those tools that we have to run.

8:40 Some of them are CPU based, some of them actually need GPUs,

8:42 some of them are simulation, some are accelerated.

8:45 Um.

8:45 On top of that, we have fleets of CPUs

8:48 that are actually managing and watching for security and governance.

8:51 We don't want our agents hallucinating, coming up with the wrong ideas,

8:56 wasting time, wasting token revenue money because

8:58 that will actually cause things to tank.

9:00 Uh, then all of that needs to be managed across that memory context window,

9:04 which is the the the working state.

9:07 The working knowledge of the model has exploded as a result.

9:11 So that's what Vera Rubin was, was designed.

9:15 And when we launched it, we didn't talk about launching a single chip

9:19 or reducing the cost of a just one LM model.

9:22 We have to actually be able to run the entire workload on the left.

9:27 Of course, the foundational platform of the Vera Rubin MVL 72,

9:31 which built on top of what we did for Grace Blackwell,

9:34 um, and expanded its capabilities.

9:36 I'll show that in a second.

9:37 We also added a booster pack.

9:39 That's the grok three.

9:41 Lf We want to go even faster.

9:43 From a user token perspective.

9:45 For those high value workloads,

9:46 we can boost up the performance of the platform even further.

9:50 Um, on top.

9:51 Now the agents CPUs matter.

9:53 We have to be able to do all that tool,

9:56 calling all that computing and be able to do it quickly,

10:00 efficiently, fully loaded.

10:01 In order to have that agent get a result quickly back to the user.

10:05 We don't want CPUs to be to ever slow down the rest of the stack.

10:09 Finally, all that KV needs to be managed in the smart stores.

10:13 We've got all sorts contexts flying around the system.

10:16 All of it needs to be managed, governed,

10:19 part of the software defined network, as well as doing indexing,

10:23 vector search and other things on the KV as we optimize it in the background

10:29 so that we have the Bluefield for STX

10:31 storage and working with all storage partners,

10:33 they build that box, but we provide them the, uh,

10:36 the smart Nic and the computing platform for them to do smart storage.

10:41 And then finally the STX networking at scale..

10:44 These models are are large that you don't fit on a single GPU.

10:50 In fact, everybody is doing disaggregated inference at this point.

10:54 We talked about it last year.

10:56 It was the Dynamo software.

10:58 It has taken off.

10:59 Um, everyone does a separate prefilled,

11:02 separate decode and all the kV on top of it that gives you the best possible

11:07 token rate and throughput and be able to connect all that you need the ESP six.

11:11 Of course, all this could be used for training as well,

11:13 and it often is in the RL and post training world.

11:17 So these are results that just got published

11:20 this weekend from the SemiAnalysis agentic benchmark on Vera Rubin.

11:24 And just threw that full stack.

11:26 Through all of that co-design with all the ecosystem,

11:30 both the open and closed models, it is actually delivering 30 x better

11:34 AI factory throughput on a genetic workloads.

11:37 In fact, in some cases up to 60 x more performance depending

11:41 on what line and what part of that Pareto curve you're looking at.

11:44 On the left hand side is the total token throughput per megawatt.

11:48 All these data centers are megawatt constrained,

11:51 and with every generation we don't think about absolute token performance.

11:56 The metric that matters is your total token throughput per megawatt.

11:59 You're increasing that metric.

12:01 You're increasing the AI factory capability to generate revenue.

12:06 If you're an operating on a model like DeepSeek v4,

12:10 you want to be you want to be thinking at a rate of over 100 tokens

12:14 per second in order to actually have it

12:16 compute an answer and then turn back around.

12:18 In fact, you won't really want to be going even further

12:21 over to the right of 200 to 50 so that the user,

12:24 the agent, can think longer, take it and come up with the answer faster.

12:29 Do those 65 turns in order to get the answer back to you,

12:33 without you having to wait a long time?

12:36 Um.

12:36 In addition, I have to also improve the the cost per million tokens.

12:41 This actually is the inverse where we actually take the.

12:45 How much does it cost to actually generate the tokens on it?

12:50 Comparing from Grace Blackwell to Vera Rubin.

12:52 As you can see here, because of the performance, the cost is reduced.

12:58 This is the biggest way we can improve.

13:01 Cost is actually deliver better performance every generation.

13:05 If I just make the GPU chip cheaper,

13:08 or I maybe optimize and use a lower cost network,

13:13 I'm only improving that one little component in this giant data center.

13:18 If I increase the performance of the of that node of that rack by ten x or 30 x,

13:27 that performance goes to the top of the equation

13:30 and the whole data center increases improves this token per cost.

13:35 So the best way to improve, uh, improve our.

13:38 We have to look at the ratio.

13:40 We have to improve performance, but look at cost.

13:43 And we do.

13:44 But the best way to actually make

13:47 tokens cheaper is actually by increasing performance.

13:49 You can.

13:50 We don't need to stop there.

13:52 What if we want to actually think faster and think even more aggressively,

13:56 and be able to do quick turns?

13:58 There are lots of workloads where agents really matter,

14:01 and we want to put them on the critical path to getting the solution and answer

14:05 for real time business analytics or computing

14:09 for and put them right at the front line.

14:13 We can boost very moving even further from Blackwell to Vera Rubin.

14:17 We increase the overall user interactivity.

14:19 But if we want to go to the extreme

14:22 where we're not talking about hundreds of tokens per second,

14:25 but thousands of tokens per second, um,

14:27 we can actually boost out the performance further to the right.

14:30 And that's using the Groq.

14:32 So These are numbers that are hot off the presses.

14:35 Um, we talked about one of them at Hot Chip's

14:38 last week and talking a new one today that the actually,

14:42 for large context, 100 K, um, uh, models either the Gemma four or the Quin 3.8.

14:50 We can actually deliver 3400 tokens per second on Gemma four

14:56 and on Quin 3.3.8 2020 529 tokens per second per user.

15:02 That's wickedly fast.

15:04 And it's on large contract context.

15:06 You know, that's the the the larger the context,

15:09 the more the model has to look over the entire history

15:13 of the conversation in order to produce every new token that increases compute.

15:17 And but it's required for generic use cases.

15:21 And then secondly the tools.

15:25 So tool calling is a critically important we have to make.

15:28 And the world is uh, not just short of their data centers,

15:32 their GPUs, but also of CPUs.

15:35 And.

15:35 While people were using the Grace attacks

15:38 attached to the Blackwell for tool calling, um,

15:41 we're now seeing the need for actually a fast,

15:45 efficient and quick CPU for agenda computing.

15:49 We've taken the CPU that was attached to the Rubin GPU,

15:53 and built a dedicated Vera Rubin rack and brought to market.

15:57 What is Vera?

15:59 It's actually our NVIDIA's own design core called Olympus.

16:04 It's it is designed to be fast, fully loaded, every core computing,

16:10 but also be able to produce the answer in the quickest possible time.

16:15 Uh, we do that through that custom Olympus core.

16:18 And there's a great white paper on if you want to read it.

16:23 Excuse me.

16:23 To make all those cores run fast and you have,

16:27 uh, an extremely high quarter core your bandwidth.

16:30 So there's no triple attacks.

16:32 This is a big monolithic compute die.

16:34 That actually be able to access.

16:37 So every core can access its memory

16:39 without having to go through different triplet.

16:43 Hops in order to get the lowest possible latency to the memory.

16:46 And you see that with with the LP DDR memory subsystem.

16:49 Not only is it low power,

16:51 but it's actually 40% lower latency than than anything else out there

16:55 with CPU calculations and math and operating

16:58 memory latency dominates often and otherwise, you'd probably be using a GPU.

17:03 So in order to build a low latency chip,

17:05 we have that the no triple attacks a single design.

17:09 But what does this show up?

17:11 How does this materialize?

17:13 Um, and it's just different than the pre cloud CPUs

17:16 where you had you had that high single threaded performance.

17:20 We didn't have a lot of cause.

17:21 So you had kind of low throughput but really fast.

17:24 This is like a gaming PC.

17:25 Um, but with only 12 cores, you had to build out quite a few

17:29 servers in order to get your total compute capability.

17:32 Um, in the world of cloud pre AI.

17:36 We basically the metric was dollar per core dollar per core dollar per core.

17:41 So the that that drove the core counts up performance was secondary.

17:45 You know don't make it any slower.

17:48 But you know improve my efficiency.

17:49 So I can deploy more cores with the same box.

17:52 And that's why you see the explosion of core counts.

17:54 Because what is the first thing people want to know when

17:57 they rent stuff from the cloud is what's my dollar per core?

18:00 Not what's my dollar per core per flop or my per per watt per dollar.

18:04 It's just dollar per core.

18:06 Because most of those tasks were, you know,

18:08 web hosting and database and other stuff that allowed businesses to scale,

18:13 um, various different it's designed to actually

18:16 deliver high throughput to every core being humming.

18:19 We have 188 cores, um, or 176 if you turn on SMP,

18:25 uh, But also deliver that high single thread performance.

18:28 Not as many courses as the rest of the cloud CPUs,

18:32 but we didn't want to sacrifice the time of those that token latency.

18:36 These are results that got published over the weekend from signal 65.

18:41 So if you go and check out the Signal65 article where they validated,

18:46 we gave them an average of C2 server

18:50 and they actually ran a bunch of CPU workloads.

18:53 These are results from a benchmark called Terminal Bench.

18:57 There's basically two big benchmarks that people

18:59 use for benchmarking SWE and Terminal Bench.

19:02 Terminal bench is from Stanford.

19:03 If anyone from Stanford here.

19:05 Thank you.

19:05 Um, but what we did here is we took actually the CPU

19:09 side of Terminal bench and just ran those agentic tests they have over,

19:14 uh, leave upwards of 80 different tests.

19:16 And so we just ran all the CPU side jobs on them.

19:21 And you can see they actually delivered 1.6 x faster performance.

19:24 The agent actually complete those tool calls 60% faster,

19:28 and some cases up to two X factor

19:32 for some of those key agent groups like compiling code,

19:36 running git, doing things like, um, code validation.

19:39 A lot of the coding benchmarks upwards of two x faster time to solution..

19:45 Really important is how fast your agent is going to work.

19:48 Barrett is is designed to run that.

19:50 And we've shown a single 65 showed that in that article.

19:54 Please do check it out.

19:55 Um, some customers that are already going and looking to deploy.

19:59 Um, one is Perplexity space.

20:01 Uh, for the running tasks,

20:03 you're often limited by how fast you can stand up a sandbox,

20:07 and by working with them and actually getting them on to Vera,

20:11 they immediately saw a like 1.5 x improvement

20:13 and how fast they could fire up a sandbox,

20:16 do a quick calculation and shut it down.

20:18 Actually, the sandbox creation dominates you on short tasks.

20:21 And then of course, for long running tests.

20:23 That's what Signal65 showed.

20:24 You also are getting that 1.5 x or up to two x faster.

20:28 These are their results.

20:29 The for their perplexity space um uh workload.

20:33 Um the other one is Clickhouse.

20:36 Clickhouse is a real time database.

20:38 It's used by all the frontier labs.

20:41 It's often used for that security and governance.

20:43 So it's it's looking at the data lake looking for issues,

20:46 looking for security things and tracking the the execution and the operations

20:51 across the and identifying issues in real time is one of the fastest,

20:57 most popular open source analytical databases out there, um,

21:00 and used by many of the data intensive companies and Frontier Labs.

21:04 Um, actually, they ran their benchmarks, their, their test suite.

21:06 This is actually an open test suite.

21:08 If you go to the clubhouse benchmark online, you can see these results.

21:12 And there is at the top of that list as a result,

21:15 they're very excited about this.

21:17 And uh, one of the key use cases is if it's going

21:20 to be for that security and governance in order to deliver good performance.

21:25 It is shipping, so definitely ask around for it.

21:29 Some of the early adopters,

21:30 which you've we've talked about and talked about before,

21:33 it's coming to all the um,

21:35 some of the, the NCP partners and other clouds, particularly uh, Oracle Cloud.

21:40 Uh, and uh, there's lots of system partners if you want to build your own.

21:45 Finally, I just want to end on a few of the notes.

21:49 Uh, NVLink is a big part of what makes AI and inference fast.

21:53 Um, you know, these are some of the.

21:55 If you just look to the left of going

21:58 from no NVLink to eight way NVLink to the full 72.

22:02 You can see how the Pareto curve improves and improves.

22:05 Improves.

22:05 And it's a complicated system level stack of building a good node architecture,

22:10 building a spine, building the multi-track and the in that entire ecosystem.

22:15 NVIDIA is a networking company and we're making this available to all.

22:19 Just like we made a spectrum available and open to to the market,

22:24 whether you're using NVIDIA GPUs or not.

22:26 We're also making every link available.

22:28 And then we do that through a program called theNVLink Fusion,

22:30 where the chip designers just need to design their accelerator, their XPU,

22:36 and partnering with NVIDIA, they can interface into with our enabling chiplet,

22:40 IP and also the stuff we've done for improving HBM.

22:44 And then at that point,

22:46 they can tap immediately into the entire NVLink ecosystem.

22:50 And all those partners on that inference needs scale up,

22:54 and we're here to help with all of that scale up, um, in fact,

22:58 and we'll open and provide access to the switch

23:00 tray and the software and the capabilities,

23:02 and they can bring all that to market.

23:05 We have a couple of early partners.

23:07 The matrix is a startup actually is doing NVLink Fusion.

23:10 Uh, this is for the Raptor next generation Raptor XPU, which will include, uh,

23:16 connecting their debut to Vera and the out

23:19 the back that actually connect to the NVLink Switch Tray.

23:22 Be able to build up a full NVLink 72 architecture with with The Matrix,

23:27 and looking forward to some of their customer announcements coming up.

23:32 And then Amazon actually and on the Annapurna

23:35 Labs is actually partnering with us.

23:37 We've already made announcements about how the next

23:39 generation training will be using NVLink fusion for scaling,

23:42 scaling up their connectivity.

23:44 And they've been a great partner.

23:47 And more recently, we've announced that they're actually

23:49 going to be partnering with us on NVHBM here.

23:52 We're actually putting custom series technology into the base die of the NVHBM.

23:57 They have to connect to their to the to the zebu on the other side.

24:02 As a result you can move all of that HBM logic and memory management

24:06 and surges to the base die frees up more compute on the expo itself.

24:12 Up to 40% more, 30% more compute.

24:15 Uh, and uh, obviously gets increased bandwidth and, uh

24:19 lower latencies because you're taking advantage of NVIDIA surges.

24:23 Um, which is, uh,, quite fast.

24:27 Okay.

24:27 Um, a little bit over time,

24:29 but I want to talk about one more thing, which is at data center scale.

24:33 We've talked a little bit about the different racks.

24:35 What about at the data center scale, one of the areas that's really, um,

24:39 as we're focused and dedicated to improving

24:42 inference per megawatt or per gigawatt,

24:44 is how much, how many GPUs can you fit inside of a data center?

24:48 And when you look at that Pareto curve,

24:52 the performance profile that trade off between,

24:54 um, uh, high power consumption at the rack level,

24:58 which tends to be on the leftmost side to the fastest token rate on the right,

25:03 changes over time, and every workload has a different performance profile.

25:06 Um, some in the early days, if you take sort of the max power,

25:10 this is just the max power of the rack.

25:11 You know, you have this many megawatts and you just roll it out.

25:14 But in reality, if you look even at a deep sea gar one,

25:17 the original DeepSeek model,

25:19 you know the power at the GPU level can vary between 600W all way up to 1900W.

25:25 Um, so the workload profile is actually varies per,

25:30 um, per rack, per use, per year.

25:32 And that that gap there is opportunity.

25:35 You're not you're provisioned for all that power but you're not using it.

25:39 In fact, you could put more GPUs into it.

25:41 And some people actually do do that.

25:43 They actually take the, uh, you know,

25:44 they don't take the max power of the rack and divide by their megawatts.

25:48 They pick a more efficient point, like a max Q point.

25:51 And so they over, they call it over provisioning by maybe 20%.

25:54 Here I've got another 20% more racks, 4800.

25:57 Uh, I've increased my total tokens per second.

26:00 But, you know, some of those workloads are going to get capped a little bit.

26:05 Uh, or, and we're not fully taking advantage of it.

26:08 So that's what our, um,

26:09 we've worked on software with the ecosystem to build something we call MaxLPS.,

26:14 Max Land Power andShell.

26:16 And what that is is basically software that can

26:19 dynamically manage the the power of every single rack.

26:22 Make sure it never exceeds a certain rack and dynamically

26:25 change the rack power up and down as the workload changes.

26:29 That gets coordinated with the data center manager,

26:32 so that we can actually deploy even our confidence to put even more racks

26:37 into that fixed data center to maximize the token revenue of the AI factory.

26:41 Uh, it's software we got to do at the firmware level.

26:44 It's a capabilities that we have all the way down to Vera

26:48 and Rubin and MP switch layer to dynamically manage and cap the power.

26:52 So we're using every last provisioned watt for that workload.

26:56 This actually can give you 40% more compute in the data center

27:00 by deploying more racks in 5600 racks in IT and a gigawatt data center,

27:05 increasing my token revenue now up to 1.2 billion tokens per second,

27:09 and this is all within the same provision power.

27:12 We are in a power limited world.

27:15 Um, if you should, the the customers and off takers

27:19 of this MaxLPS need to be asking for MaxLPS capability so that they

27:23 feel so that we can actually make sure we're offering

27:26 and the bringing the most possible GPUs and most token revenue to market.

27:31 Um, we have some early partners on this.

27:34 Lambda actually has did a great job of, uh,

27:36 benchmarking and testing and validating the stack.

27:39 And you should go check out there.

27:41 Dave Ward um, uh, is very excited about it.

27:44 He's actually able to increase the number of GPUs,

27:46 number tokens he can offer to his to his customers upwards of, uh, on the B200.

27:51 It was it was more like 25% on Vera Rubin it's

27:54 going to be upwards of 40% more compute in the same megawatts.

27:58 And it's coming to a lot of the NCP partners and working with the CSPs as well.

28:03 And that's it.

28:04 So that's what DSX MaxLPsis.

28:06 It's about optimizing our software stack and building

28:08 a reference architecture for debuting data center at scale.

28:12 Maximize the revenue.

28:13 Maximize the number of GPUs.

28:15 Maximize the infrastructure and operating system

28:17 because you have to at that scale

28:19 in order to make things successful and get the best possible infrastructure,

28:23 built out infrastructure that's fungible, that's reliable,

28:26 that's going to last ten years and really becomes an asset

28:31 for computing the world's computing problems in the world's industries.

28:35 They can build out with Vera Rubin,

28:37 with DSX and all of the innovations that we have here.

28:40 NVIDIA is committed to a one year roadmap.

28:43 This is Vera Rubin here.

28:44 Next time I'm on the stage, I'm going to be talking about Vera Rubin Ultra.

28:47 And after that I'm going to talk about Feynman.

28:49 Now that one's about Feynman ultra.

28:50 You can trust that NVIDIA is going to be is on an annual cadence.

28:54 And we're continually dedicated to advancing, uh,

28:58 the AI token factory advancing inference and agents at data center scale.

29:03 Thank you very much and appreciate the time.